This repo provides a reference implementation of Risk-ConversationGraph as described in the paper Reading Between the Lines: A Graph-based Financial Risk Prediction Approach:

While the earnings transcript dataset cannot be publicly released due to the data provider's policy, we make our code implementation publicly available.

We hope that our design can benefit researchers and practitioners and shed light on other financial prediction tasks

Note: Researchers can easily obtain the earnings conference call data from Seekingalpha or databases such as Thomson Reuters StreetEvents.

Run the following to install a subset of necessary python packages for our code

pip install -r requirements.txt

Train and evaluate our models through main.py. Here are some vital options:

main.py:
  --config: Training configuration. Readers can change it to adapt to their specific tasks.
    (default: 'conversation_graph.yml')
  --trainer: Init and run model. In our design, we bound trainer with their specific model.
    (default: 'ConversationGraph')
  --test: Whether to evaluate the model with the best checkpoint.

After training, turn on --test to test model with best checkpoint. We select the best checkpoint according to the MSE results on the validation dataset.

• assets: contains best_checkpoint.pth and lda_topic_rep.pkl. Necessary for computing MSE, MAE, Spearsman's rho and Kendal's tau.
• data: contains data_2015.pkl, data_2016.pkl, data_2017.pkl, data_2018.pkl that represent data of different years. The detailed content of each .pkl can be found in trainer/conversation_graph_trainer.py -- loading_train_dataset function.
• model: contains conversation_graph.py. Our model's file.
• utils: We implement attention and other functions here.

We use pre-trained Bert-based-un-cased to generate sentence embeddings. Note that other pretrained language models such as RoBERTa, SentenceBert Sentence Bert, and FinBert can also be used as the text encoder, but we find the results are similar.

We select the topic amounts of the LDA model by coherence score:

import os
import gensim
from gensim.models import CoherenceModel
def compute_coherence_values(dictionary, corpus, texts, limit, start=2, step=3):
    """
    Compute c_v coherence for various number of topics
    Parameters:
    ----------
    dictionary : Gensim dictionary
    corpus : Gensim corpus
    texts : List of input texts
    limit : Max num of topics
    Returns:
    -------
    model_list : List of LDA topic models
    coherence_values : Coherence values corresponding to the LDA model with respective number of topics
    """
    coherence_values = []
    model_list = []
    mallet_path = 'packages/mallet-2.0.8/bin/mallet'
    for num_topics in range(start, limit, step):
        model = gensim.models.wrappers.LdaMallet(mallet_path,
                                                 corpus=corpus,
                                                 num_topics=num_topics,
                                                 id2word=dictionary,
                                                 random_seed=1234,
                                                 workers=os.cpu_count())
        # model_list.append(model)
        coherencemodel = CoherenceModel(model=model, texts=texts, dictionary=dictionary, coherence='c_v')
        coherence_values.append(coherencemodel.get_coherence())

    return model_list, coherence_values

Some NLP preprocessing techniques are needed, including converting text to lowercase, removing emojis, expanding contractions, removing punctuation, removing numbers, removing stopwords, lemmatization, etc.